Training a world model to discriminate between candidate next-states beats training it to predict them
arXiv 2609.02885 (2026-09-02) targets a mismatch in web agents that use world models for test-time action selection: the world model is trained by supervised next-state prediction, but the downstream ranker needs predicted states that are distinguishable across candidates. Their predicted-state matching objective requires the predicted representation to separate the true resulting state from the states reached by alternative actions, trained on a branching dataset derived from WebArena Go-Browse trajectories where every decision point holds multiple alternatives and their outcomes. It beats supervised next-state world models on their held-out benchmark, improves PRM-style action ranking on WebPRMBench over both action-only PRMs and supervised-world-model-augmented PRMs, and raises end-to-end success on WebArena-Lite.
↳ Follow the thread